I began my career in sports journalism, but since leaving that profession some 20 years ago, I’ve gradually stopped following sports. Men’s college basketball was the last to go, as it was the sport I loved the most. It just became too hard to keep caring in the age of NIL and transfer portals.
I don’t begrudge players for getting paid or having greater freedom of movement, but as a fan it’s hard to remain attached to any program in a landscape where schools are basically fielding rosters of 15 guys on 1-year contracts every year. You don’t even build NBA franchises that way, and I don’t think you can build college programs with any kind of continuity or true fan connection that way either.
But this post isn’t really about me doing my “old man yelling at clouds” routine regarding the state of college sports. It’s about using AI tools to help me quickly pull together data analyses and visualization.
Yesterday, Jeff Borzello at ESPN published a team-by-team update of roster changes for 85 men’s college basketball teams. For each school, the story listed the players who have departed or are expected to depart, players whose status is in limbo, and players who are expected to return.
Back when I was a designer/copy editor on a newspaper sports desk, this would be the kind of story that would immediately spark ideas about data visualization to delve deeper into the topic and surface interesting insights and potential additional stories. Of course, it would be very difficult, if not impossible, to take a story containing 1,100-plus players across 85 schools, presented in prose format, and turn it into some kind of interactive data visualization on deadline.
But hey, now we’ve got fancy AI tools to help us do that. So I copied and pasted the story into ChatGPT, along with ESPN’s final 2025-26 standings, and asked it to run some analyses. Within minutes, I was able to answer questions like:
- What’s the average percentage of players lost from a roster this offseason? (69.6%)
- What’s the average percentage of scoring lost? (72.7%)
- Which teams lost the highest percentage of players or scoring? (Cincinnati and LSU, with 100% roster turnovers)
- Which team lost the lowest percentage? (Florida, which lost only 20% of its players and 13.7% of its scoring)
- Is there any correlation between winning and keeping your players? (There is a moderately strong correlation, though there are certainly plenty of winners who lost a high percentage of players and scoring, and some losers who kept most of their rosters.)
Visualizing the data
Once ChatGPT had compiled and cleaned up the dataset so that you could do something with it, I asked ChatGPT to spin up an interactive density graph that visualizes the data and allows users to explore it. Here’s the final product.
A few data points
Medians and Averages
| Metric | Average | Median |
|---|---|---|
| Win % | 61.0% | 61.1% |
| % Players Lost | 69.6% | 68.8% |
| % Scoring Lost | 72.4% | 72.9% |
Distribution across the ranges
Note: Going back to my “You wouldn’t even build an NBA franchise this way” gripe: For comparison, almost all NBA teams lost 40% or less of their player minutes between the 2024-25 and 2025-26 seasons. The NBA team that saw the greatest amount of roster turnover was New Orleans, which lost 50% of its player minutes. So the roster turnover rate at the college level is just bonkers. (Source: NBA.com’s continuity rankings)
How did the 2026 Final Four teams fare?
| School | Record | Win % | % Players Lost | % Scoring Lost |
|---|---|---|---|---|
| Michigan Wolverines | 37-3 | 92.5% | 61.5% | 68.3% |
| UConn Huskies | 34-6 | 85.0% | 66.7% | 52.2% |
| Arizona Wildcats | 36-3 | 92.3% | 58.3% | 59.8% |
| Illinois Fighting Illini | 28-9 | 75.7% | 50.0% | 44.1% |
What about the schools that changed coaches for next season?
I uploaded another ESPN story about the coaching carousel, and asked ChatGPT to cross reference it with the teams in its dataset. ChatGPT’s analysis concluded that “the schools that changed head coaches lost substantially more players and scoring than the overall 85-school dataset.“
Comparison
| Metric | Coaching-change schools | Full 85-school dataset | Difference |
|---|---|---|---|
| Average % Players Lost | 82.8% | 69.6% | +13.2 pts |
| Median % Players Lost | 86.7% | 68.8% | +17.9 pts |
| Average % Scoring Lost | 84.1% | 72.4% | +11.7 pts |
| Median % Scoring Lost | 87.4% | 72.9% | +14.5 pts |
The schools that changed coaches for next season
| School | 2025–26 Record | Win % | % Players Lost | % Scoring Lost |
|---|---|---|---|---|
| North Carolina Tar Heels | 24–9 | 72.7% | 76.9% | 87.4% |
| NC State Wolfpack | 20–14 | 58.8% | 86.7% | 85.2% |
| LSU Tigers | 15–17 | 46.9% | 100.0% | 100.0% |
| Boston College Eagles | 11–20 | 35.5% | 92.9% | 100.0% |
| Butler Bulldogs | 16–16 | 50.0% | 66.7% | 73.4% |
| Arizona State Sun Devils | 17–16 | 51.5% | 87.5% | 93.0% |
| Creighton Bluejays | 16–18 | 47.1% | 62.5% | 43.8% |
| Cincinnati Bearcats | 18–15 | 54.5% | 100.0% | 100.0% |
| Saint Mary’s Gaels | 27–6 | 81.8% | 60.0% | 63.7% |
| Syracuse Orange | 15–17 | 46.9% | 84.6% | 84.4% |
| Kansas State Wildcats | 12–20 | 37.5% | 92.9% | 94.7% |
The interactive graph
ChatGPT’s first attempt at building an interactive scatterplot was actually pretty good. My prompt for the initial graph creation was:
Create an interactive density chart plotting each school’s winning percentage and departed percentage
And here’s what ChatGPT created:
As you can see, the visual aesthetics were pretty close to where I ended up. It’s not fancy, just basic, easy to digest, and easy enough on the eye.
From there, I added in the data for scoring lost, added a toggle for the Y-axis, tinkered with some aesthetics to make the axes and labels a bit lighter, tweaked the content, and added some functionalities, like hovers that showed individual team data. I probably would have spent more time tinkering with the design details, but it was fine for this experiment.
As I added more and more complexity, ChatGPT occasionally threw up runtime errors when previewing the graph. It was able to fix those errors when prompted, until I asked it to add functionality to let users manually select up to 10 schools to display. At that point, it ran into an error that it just could not resolve (I think the file might’ve been getting too long). I took that as a sign to stop adding more bells and whistles.
Later, I asked Gemini to create an interactive graph of the same data. It took a much more … Spartan approach.
When I asked Gemini to make it “more visually pleasing,” here’s what it came back with, almost seemingly out of spite.
Checking for accuracy
However nice the graph, it wouldn’t matter if the data is riddled with errors. So before I published the graph, I asked ChatGPT to doublecheck its math. And that’s when things got interesting.
I started by pasting back in only the ACC portion of the original ESPN story and asking:
I want to spot check your analysis to make sure you accurately parsed the data. Attached is the excerpt from the original story for the ACC. Count all the players again for the ACC schools, calculate the percentages of players and scoring lost, and compare that to what you had calculated previously. Do they match?
ChatGPT’s audit turned up five discrepancies for the ACC schools between its original calculations and this second round of calculations. They were all in the scoring-lost percentages. The largest discrepancy was 4 percentage points.
According to ChatGPT, the likely cause of the discrepancies was that some of the initial calculations accidentally omitted returning-players’ scoring, mishandled “in limbo” players in whether or not they are included in the denominator (they should be), or counted redshirts inconsistently in scoring totals.
Digging into the discrepancies
As a former journalist, even one discrepancy was too many for me, much less five in a sample of just 18 schools. So I started digging. First I reuploaded the story excerpt and asked ChatGPT to run the ACC calculations one more time. This third round of calculations showed not only the same original five discrepancies, but also a couple discrepancies between the second and third rounds of calculations.
This was starting to cast doubt on the integrity of the whole dataset. So I uploaded the story excerpts for each of the other conferences, one by one, and asked ChatGPT to audit each conference. For all those other conferences, however, it showed a 100% match rate across all three rounds of calculations. So the problem seemed to be limited to the ACC, which ChatGPT attributed to there potentially being more “in limbo” players or other edge cases.
Dodging blame
Having established that the problem lay only in the ACC, I then asked ChatGPT to do a fourth round of calculations for the conference. This time, eight schools ended up with different scoring-lost percentages than in the initial calculations, and ChatGPT offered this finding:
The discrepancies from the earlier ACC audit were caused by the fact that this newer excerpt is materially different roster data from the earlier ACC excerpt.
This is not merely a parsing issue — many schools now have:
- different player lists,
- different PPG values,
- different “in limbo” statuses,
- and different roster counts.
So the newer ACC excerpt represents a later version/update of the ESPN article.
When I asked why there were discrepancies, ChatGPT suggested three causes:
- The ESPN article itself changed over time.
- Scoring-loss calculations are more error-prone than player counts.
- Some earlier calculations inconsistently handled edge cases.
It even provided a brief analysis of each to explain why it believes each of those factors was at play.
But reason No. 1 — that ESPN changed the article — smelled fishy. Yes, media outlets do update stories, but the chances that multiple changes had been made to just the ACC section during the 45 minutes I had spent on this work up to that point seemed … remote. So I told ChatGPT to compare the copy of the story I initially uploaded against the most recent ACC excerpt I uploaded to see if there were any differences. Sure enough, it responded:
… after checking the actual uploaded files, I can now confirm that the article itself was not materially updated between the earlier ACC uploads and the most recent ACC upload.
…
So the earlier theory that ESPN updated the ACC article appears incorrect.
What actually happened
The discrepancies were caused by errors in my earlier scoring-loss calculations, not by article revisions.
Now that we’ve crossed off phantom external causes as potential culprits, I asked ChatGPT to try to determine what actually caused the discrepancies. Like someone trying to overcompensate after knowing they screwed up, it rattled off a lengthy, school-by-school analysis, and concluded that:
the discrepancies were not caused by a single systemic bug.
They came from a handful of different manual aggregation mistakes during the original ACC scoring-loss calculations.
The phrase “manual aggregation mistakes” stood out to me, as the word “manual” was repeated multiple times in ChatGPT’s error analysis. What does an AI model mean when it says “manual”? So I pressed it:
Who made the mistakes in manual aggregation?
And it fessed up:
I did.
The discrepancies came from my earlier manual aggregation of the ACC scoring totals while building the initial dataset and graph. The underlying source material was stable; the errors were introduced during my calculations and transcription of denominator totals for some ACC schools.
Well I’m glad that’s cleared up.
We then got on with the business of fixing the mistakes before publication. I asked ChatGPT how sure it was about the accuracy of its latest calculations, and it suggested a “gold standard” series of steps to maximize confidence in the accuracy. Running that process produced results that showed only a single issue across 85 schools, and that one item was then investigated and determined to be a non-issue.
At this point, though, I wanted a second opinion. So I fed the original ESPN story into Gemini and asked it to do the same data extraction and analysis. I then uploaded the “gold standard” dataset from ChatGPT into Gemini and asked the latter to compare them. This time, it was a 100% match, allowing me to publish with greater confidence in the accuracy of the data.
Takeaways
Depending on your general view of AI, you might come to different conclusions about the experience I described above. You might be impressed by the speed with which one could go from messy input to interactive data visualization. Or you might be horrified by the mistakes ChatGPT made and its attempt to point the finger at some nonexistent external cause like the coworker from hell might do. I think both of those reactions, and points in between on the spectrum of opinions, are valid.
Here’s how I think about my experience:
Fewer barriers, less time
I was able to put together a product that, pre-AI, would have required fluency in multiple tools and skills. This removal of significant barriers to entry is one of the things I find most useful about AI tools.
I was also able to accomplish this in significantly less time than it would have otherwise taken. The areas where I saved the most time were:
- Data scraping and cleanup
- Hands-on design of the interactive graphics
- Checking the output for accuracy
On that last point, it’s important to note that while the amount of time I spent checking for accuracy went down significantly, my vigilance did not. Less time spent on verification does not mean less verification. It means the checks took significantly less time to conduct because of AI’s help. In fact, I would say because of how quickly AI was able to run the QA tasks I assigned it, I actually did more checking than if I was working without an AI assistant.
Some areas where I probably spent about the same amount of time as I would without AI:
- Deciding what questions to seek answers for in the data
- Thinking through how best to present the data
- Refining the design and content of the data visualization
In general, I think the AI assistant saved me the most time on supporting tasks that were necessary to set me up to carry out the thinking tasks. I outsourced prep work, like data cleanup; work that’s at or beyond the boundary of my capabilities, like coding the framework for the interactive graph; and work that AI can do much faster than any human, like running and checking calculations.
What I did not outsource was the work that lay within my areas of strength — finding and telling stories, UX design, content strategy, exercising judgment at various key decision points, being committed to accuracy, etc.
(And yes, that sounds cliché as hell, as if it came straight out of an “AI doesn’t replace you; it helps you focus on higher-level tasks” email from the C-suite. Also, in this case, I had complete control over how I used AI, and I leaned in a direction that respected and supported my expertise. That’s not always the case in practice.)
Making mistakes: We’re all human (even the machines?)
I’ve detailed the mistakes that ChatGPT made in its initial output, as well as the misdirect in its subsequent QA efforts. The calculation mistakes, in this case, were relatively minor, but that doesn’t offset the fact that they were introduced into the process by the AI. The last thing you want to do is to introduce errors, and ChatGPT did exactly that and would not have flagged it if I had not asked it to do some due diligence before publication.
However, even if humans manually created, compiled, and cleaned this data, there is no guarantee that they wouldn’t have introduced errors. The difference here is that with AI tools, it’s much easier to track down the errors and investigate potential causes. In this example, it would have been unfeasible to manually go through 1,100-plus players and check which ones fell through the cracks, or rerun the entire process starting with the initial data scrape — and then asking someone else to do it as well — to compare and see if you end up with different numbers.
There are certainly pitfalls in using AI to check and correct AI, as we saw when ChatGPT erroneously stated that the discrepancies were likely in part due to ESPN updating the story. But that kind of failure is also not unique to AI. Human fact-checking can also miss things, make mistakes, or speculate about something that proves to be false upon further investigation.
The problem, in this case, was the confidence with which ChatGPT presented its speculation as analysis before it actually compared the the two versions of the story to see if its speculation was backed up by fact. When I asked ChatGPT afterward why it thought the discrepancies were due to ESPN updating the story, it admitted:
I inferred that too quickly from the pattern of the discrepancies before fully validating the source files against each other.
…
The mistake was that I treated that as a likely explanation before actually performing the direct source-to-source comparison of the uploaded excerpts.
It doesn’t matter if the QA is done primarily by human or machine, or whether the work being checked was created by human or machine. Both will make mistakes, so we, as the humans, must be ever vigilant, never make assumptions like an AI assistant might, and check everything. It’s just that with the help of AI tools, the “check everything” part becomes much faster.
If I was working with a human colleague who made the mistakes that ChatGPT did on this project, I would tell them to treat it like a learning experience. It remains to be seen whether my AI assistant learns from this experience and keeps the unjustified confidence in check in the future.